Message-ID: <20070104141638.GB11203@kernel.dk>
Date:	Thu, 4 Jan 2007 15:16:38 +0100
From:	Jens Axboe <jens.axboe@...cle.com>
To:	saeed bishara <saeed.bishara@...il.com>
Cc:	linux-kernel@...r.kernel.org
Subject: Re: using splice/vmsplice to improve file receive performance

On Thu, Jan 04 2007, Jens Axboe wrote:
> On Wed, Jan 03 2007, saeed bishara wrote:
> > On 12/22/06, Jens Axboe <jens.axboe@...cle.com> wrote:
> > >On Fri, Dec 22 2006, saeed bishara wrote:
> > >> On 12/22/06, Jens Axboe <jens.axboe@...cle.com> wrote:
> > >> >On Fri, Dec 22 2006, saeed bishara wrote:
> > >> >> On 12/22/06, Jens Axboe <jens.axboe@...cle.com> wrote:
> > >> >> >On Thu, Dec 21 2006, saeed bishara wrote:
> > >> >> >> Hi,
> > >> >> >> I'm trying to use the splice/vmsplice system calls to improve the
> > >> >> >> samba server write throughput, but before touching the smbd, I 
> > >started
> > >> >> >> to improve the ttcp tool since it simple and has the same flow. I'm
> > >> >> >> expecting to avoid the "copy_from_user" path when using those
> > >> >> >> syscalls.
> > >> >> >> so far, I couldn't make any improvement, actually the throughput 
> > >get
> > >> >> >> worst. the new receive flow looks like this (code also attached):
> > >> >> >> 1. read tcp packet (64 pages) to page aligned buffer.
> > >> >> >> 2. vmsplice the buffer to pipe with SPLICE_F_MOVE.
> > >> >> >> 3. splice the pipe to the file, also with SPLICE_F_MOVE.
> > >> >> >>
> > >> >> >> the strace shows that the splice takes a lot of time. also when
> > >> >> >> profiling the kernel, I found that the memcpy() called to often !!
> > >> >> >
> > >> >> >(didn't see this until now, axboe@...e.de doesn't work anymore)
> > >> >> >
> > >> >> >I'm assuming that you mean you vmsplice with SPLICE_F_GIFT, to hand
> > >> >> >ownership of the pages to the kernel (in which case SPLICE_F_MOVE 
> > >will
> > >> >> >work, otherwise you get a copy)? If not, that'll surely cost you a 
> > >data
> > >> >> >copy
> > >> >>   I'll try the vmplice with SPLICE_F_GIFT and splice with MOVE. btw,
> > >> >> I noticed that the  splice system call takes the bulk of the time,
> > >> >> does it mean anything?
> > >> >
> > >> >Hard to say without seeing some numbers :-)
> > >> I'm out of the office, I'll send it later. btw, my test bed ( the
> > >> receiver side ) is arm9. does it matter?
> > >
> > >The vmsplice is basically vm intensive, so it could matter.
> > >
> > >> >> >This sounds remarkably like a recent thread on lkml, you may want to
> > >> >> >read up on that. Basically using splice for network receive is a bit 
> > >of
> > >> >> >a work-around now, since you do need the one copy and then vmsplice 
> > >that
> > >> >> >into a pipe. To realize the full potential of splice, we first need
> > >> >> >socket receive support so you can skip that step (splice from socket 
> > >to
> > >> >> >pipe, splice pipe to file).
> > >> >> Ashwini Kulkarni posted patches that implements that, see
> > >> >> http://lkml.org/lkml/2006/9/20/272 .  is that right?
> > >> >> >
> > >> >> >There was no test code attached, btw.
> > >> >> sorry, here it is.
> > >> >> can you please add sample application to your test tools (splice,fio
> > >> >> ,,) that demonstrates my flow; socket to file using read & vmsplice?
> > >> >
> > >> >I didn't add such an example, since I had hoped that we would have
> > >> >splice from socket support sooner rather than later. But I can do so, of
> > >> >course.
> > >> do you any preliminary patches? I can start playing with it.
> > >
> > >I don't, Intel posted a set of patches a few months ago though. I didn't
> > >have time to look that at the time being, but you should be able to find
> > >them in the archives.
> > >
> > >> >I'll try your test. One thing that sticks out initially is that you
> > >> >should be using full pages, the splice pipe will not merge page
> > >> >segments. So don't use a buflen less than the page size.
> > >>
> > >> yes, actually I  run the ttcp with -l65536 ( 64KB ), and the buffer is
> > >> always page aligned.also, the splice/vmsplice with MOVE or GIFT will
> > >> fail if the buffer is not a whole pages. am I rigth?
> > >
> > >Yes.
> > >
> > >I added a simple splice-fromnet example in the splice git repo, see if
> > >you can repeat your results with that. Doing:
> > >
> > ># ./splice-fromnet -g 2001 | ./splice-out -m /dev/null
> > >
> > >and
> > >
> > ># cat /dev/zero | netcat localhost 2001
> > >
> > >gets me about 490MiB/sec, using a recv/write loop is around 413MiB/sec.
> > >Not migrating pages gets me around 422MiB/sec.
> > >
> > >--
> > >Jens Axboe
> > >
> > >
> > I've done some investigation in the splice flow and found the following:
> > even when using vmsplice with GIFT and splice with MOVE, the user
> > buffers still copied, I see that the memcpy from pipe_to_file() is
> > called.
> > I added debug messages in this function and here what I got:
> > 1. the  generic_pipe_buf_steal always fails, this is because the
> > page_count is 2.
> > 2. after then, the find_lock_page fails as well.
> > 3. page_cache_alloc_cold succeeds.
> > 4. but, since the buf->page is differs from the page (returned by
> > page_cache_alloc_cold) the memcpy function is called.
> > 
> > this behavior true for all the buffers that vmspliced to ext3 file.
> > is this the expected behavior? is there any way to make the steal
> > operation return with success?
> 
> It works for me, with most pages. Using the vmsplice/splice-out from the
> splice tools, doing
> 
> $ ./vmsplice -g |  ./splice-out -m g
> 
> about half of the pages have count==1 and the steal suceeds.
> 
> find_lock_page() will only suceed, if the file exists and is cached
> already. splice-out will truncate the file, so it should never suceed
> for that case. For both the find_lock_page() success and failure case
> (page being allocated), it's a given that we need to copy the data.

Testing a simpler case (not switching buffers), all but one page was
stolen. I tested with on-stack and posix_memalign returned buffers.

-- 
Jens Axboe

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists